ppae

Training feedbacks on the dataset of 1000 eventfiles

current problems

• Some data summary:

  • properties_full.csv:
    • total length: 130232
    • with unique (region_id, obsid): 129527
  • eventfiles_full.csv

    • with unique (region_id, obsid): 189389

    • with those also appearing in properties_full.csv: 95821

    • Harness ratio

    • Variability
  • Different datasets:
    • large_filtered: 14891
    • large: 109656

• Interesting sources to look at:

  • 2CXO J011908.6-341130 (obsid 22096 region ID 9).
    • No info?
  • 2CXO J005440.5-374320 16029 162
    • [64709, 64710]
  • 2CXO J175046.8-311629 21218 105
    • No info
  • 2CXO J140314.3+541806 4732 10
    • [1792, 1793]
  • 2CXO J132943.3+471134 13814 567
    • [90436, 90437, 90438, 90439, 90440, 90446]

• Questions:

  • What other quantities to look at
  • What I already wrote
  • What else plot to put there

What did not work

• Small dataset

  • Include time difference in encoding
  • Encoder
  • Mixed precision training REALLY HURTS!

• Large dataset

  • accumulate gradient doesn't help a huge

What did work

• General:

  • TV loss on usual (exp) rates
  • TV should be separate: event rate + mesh rate, this is a lot faster
  • Plotting should use total rate
  • Train longer
  • For positional encoding, should normalize T to be 1, and include t
  • include random shift at 0 to prevent overfitting

• Large dataset

  • Potentially eliminating nan/inf values
    • Skip nan loss gradient update (done)
    • torch anomaly detection
    • Gradient clipping
    • Initialization
    • Batch normalization
  • Balance dataset

ToDO list

• visualize classes in tsne
• interesting event files
• change reference chandra
• YSO vs AGN classification